In this article (8 sections)
The AI Act’s high-risk rules have been postponed, but the Article 50 transparency obligations have not. They have applied since 2 August 2026, and for generative AI systems placed on the market before 2 August 2026, the grace period for machine-readable marking runs to 2 December 2026. Website chatbots, AI customer service, AI-generated product images and promotional videos can all fall under this article. This article explains who is responsible for each of the four Article 50 obligations, and where Taiwanese companies most often get it wrong.
Key takeaways
- Article 50 contains four obligations: informing people they are interacting with a chatbot (provider), machine-readable marking of generated content (provider), informing people about emotion recognition and biometric categorisation (deployer), and disclosing deep fakes and text on matters of public interest (deployer).
- Companies that use off-the-shelf AI tools to create marketing materials are usually deployers; but if you package a model as your own branded chatbot on your website, you may become a provider.
- AI-generated images, audio and video that “resemble” existing persons, objects, places or events, and would falsely appear to be authentic, are deep fakes, and the deployer must disclose them. Product images are no exception.
- Breaches of Article 50 carry maximum fines of EUR 15 million or 3% of worldwide annual turnover, whichever is higher (Article 99(4)(g)).
What are the four Article 50 obligations?
| Paragraph | Obligation | Who is responsible | Exception |
|---|---|---|---|
| Article 50(1) | AI that interacts directly with natural persons must let them know they are interacting with AI | Provider | Where this is obvious from the context |
| Article 50(2) | AI that generates synthetic audio, images, video or text must mark its output in a machine-readable format, detectable as AI-generated | Provider (including general-purpose AI systems) | Standard editing assistance only, without substantially altering the input |
| Article 50(3) | Deployers of emotion recognition or biometric categorisation systems must inform the persons exposed | Deployer | Use permitted by law to detect crime |
| Article 50(4) | Deep fake images, audio and video must be disclosed as AI-generated or manipulated; AI-generated text published to inform the public on matters of public interest must also be disclosed | Deployer | Evidently artistic, creative or satirical works only need appropriate disclosure; text that has undergone human review and for which someone holds editorial responsibility is exempt |
Article 50(5) requires the information to be provided in a clear and distinguishable manner at the latest at the time of the first interaction or exposure, and to meet accessibility requirements.
Also note the relationship between Article 50(3) and Article 5: using AI to infer emotions in the workplace and in education institutions, except for medical or safety reasons, is a prohibited practice under Article 5(1)(f). Informing people is not enough.
We only “use” AI tools. What are we responsible for?
That depends on your role under Article 50:
- You use off-the-shelf tools such as ChatGPT or Midjourney to create materials: the tool vendor is the provider and is responsible for the machine-readable marking under Article 50(2); you are the deployer, and your main obligation is deep fake disclosure under Article 50(4).
- You use a model API to build your own branded customer service chatbot on your website or app: by putting an AI system into service under your own name, you are very likely the provider of that chatbot and responsible for designing the notice under Article 50(1).
- You build generative features into products or software that you sell to customers: you are the provider, and both Article 50(1) and Article 50(2) apply.
Article 2(1)(c) of the AI Act provides that providers and deployers established in a third country are within scope if the AI output is used in the EU. So a chatbot on an English-language website run from your Taiwan headquarters, or catalogues and videos aimed at European customers, cannot be excluded on the grounds that “we are not in the EU”.
Does any of this apply to your company?
- Your website, app or messaging customer service uses AI auto-replies
- Product images, lifestyle photos or endorsement videos for the European market are generated or heavily edited with AI
- The software or equipment you sell has built-in generative AI features
For generative AI systems placed on the market before 2 August 2026, the Article 50(2) marking grace period runs to 2 December 2026.
See the AI Act timeline and milestones in the Regulatory Radar →
Is an AI-generated product image a “deep fake”?
Many people think deep fakes only concern politicians and celebrities. The definition in Article 3(60) of the AI Act is broader: AI-generated or manipulated image, audio or video content that resembles existing persons, objects, places, entities or events and would falsely appear to a person to be authentic.
Applied to marketing materials, three common situations deserve attention:
- AI-generated “lifestyle” photos: the product is shown in what looks like a real factory or a real home, and viewers would assume it was photographed.
- AI-modified product appearance: colours, materials or sizes not yet in mass production are rendered as if photographed.
- AI-generated endorsement or customer testimonial videos: they look like a real person speaking.
Whether these situations amount to a deep fake depends on the content and the way it is presented. In addition, Article 50(6) states that the article does not affect transparency obligations under other EU or national law. Misleading images in advertising also fall under the rules on unfair commercial practices, and environmental claims must also be checked against the new anti-greenwashing rules (ECGT).
Timeline: what already applies, and what still has a grace period?
| Date | What applies | Provision |
|---|---|---|
| 2 February 2025 | Prohibition of emotion inference in the workplace and education institutions (except for medical or safety reasons) | Article 5(1)(f), Article 113 |
| 2 August 2026 | Article 50(1) to (6) apply | Article 113 |
| 2 December 2026 | For generative AI systems placed on the market before 2 August 2026, the Article 50(2) marking grace period ends; the prohibition on generating non-consensual intimate images applies | Article 111(4), Article 5(1)(ba) |
The Digital Omnibus did not change the content of Article 50(1) to (6). It only amended Article 50(7): the Commission promotes codes of practice on labelling and detection, and if it considers a code inadequate, it can lay down common rules by implementing act. The technical level that marking must reach will be updated over time in line with codes of practice and standards.
The three most common mistakes
Mistake 1: “We wrote ‘AI-generated’ in the corner of the image, so Article 50(2) is covered.” Article 50(2) requires machine-readable marking and applies to the provider of the AI system. That is a different thing from a deployer adding a note on screen. Conversely, a deployer’s deep fake disclosure is meant for people to read. One cannot replace the other.
Mistake 2: “Our staff edited the content, so it no longer counts as AI-generated.” The exception for content that “has undergone human review and for which someone holds editorial responsibility” appears only in the second subparagraph of Article 50(4), for text on matters of public interest. There is no such exception for deep fake images, audio or video.
Mistake 3: “Our customer service chatbot uses a big-name model, so the responsibility lies with the model vendor.” The model vendor is responsible for obligations at the model level; the company that launches a chatbot under its own name is the provider of that AI system. The notice has to be designed into the interaction interface, and only the business that deploys the chatbot can do that.
How far does disclosure need to go?
When and where to inform people and what to say, how light deep fake disclosure can be for artistic work, and how to confirm machine-readable marking with your vendors: there are several approaches, and the right one depends on your channels and content types.
FAQ
Does AI-generated content always have to be labelled in the EU?
It depends on your role and the content. Providers of generative AI systems must always mark output in a machine-readable format (Article 50(2), except for standard editing assistance). Deployers must disclose to the public when content is a deep fake, or when they publish AI-generated text on matters of public interest (Article 50(4)).
When does Article 50 apply?
From 2 August 2026. For generative AI systems placed on the market before 2 August 2026, the Article 50(2) marking obligation has a grace period to 2 December 2026 (Article 111(4)).
Is it enough for our website chatbot to say “Hello, I am your smart assistant”?
Article 50(1) requires that the people concerned are informed that they are interacting with an AI system, unless this is obvious from the context. Whether “smart assistant” is enough for a reasonable person to understand they are dealing with AI depends on the overall interface design. The information must be provided at the latest at the first interaction.
What happens if we breach Article 50?
Under Article 99(4)(g), maximum fines are EUR 15 million or 3% of worldwide annual turnover, whichever is higher; for SMEs, whichever is lower (Article 99(6)). The actual penalties are set by each Member State.
Next in the series
Part 3: Are you a provider or a deployer? Determining your AI Act role
Previous: Part 1: The new AI Act timeline after the Digital Omnibus
Need help?
ZC Zeelandia Consulting has an office in Amsterdam and can help you take stock of the AI tools and content you use for Europe, and confirm AI Act applicability and your role.
- Book a 30-minute online consultation: calendly.com/zc-meeting/30min
- EU Regulatory Radar: AI Act page
- Follow ZC on LINE: lin.ee/QKV5RTE
Legal sources and references
- Regulation (EU) 2024/1689 (Artificial Intelligence Act), Article 3(60), Articles 5, 50, 99 and 111: EUR-Lex
- Regulation (EU) 2026/1744 (Digital Omnibus on AI): EUR-Lex
- AI Act consolidated text (27 July 2026): EUR-Lex
- European Commission, AI Act Service Desk FAQ: ai-act-service-desk.ec.europa.eu
This article provides general information, updated to November 2026, and does not constitute legal advice. Whether a specific product is affected depends on the actual facts and the latest Official Journal texts.
Next step: have you considered these three options?
Sources · Regulation (EU) 2024/1689 (Artificial Intelligence Act) · Regulation (EU) 2026/1744 (Digital Omnibus on AI) · AI Act consolidated text (27 July 2026) · European Commission, AI Act Service Desk FAQ. This article is general information and not legal advice for any specific case. Obligations are defined by the Official Journal of the EU.
AI ActArticle 50transparency obligationsgenerative AIAI-generated content labellingdeep fakechatbot2024/1689
